Four task-built agents.
One claims workflow.

A first notice of loss arrives in whatever shape the person reporting it had to hand: a call from a supervisor at the end of a shift, an email with a photographed form attached, a broker's PDF, a voicemail left after hours. The intake agent takes all of them, in the channel they arrive on.

It reads or listens, extracts the fields your schema requires, checks them against the policy and the parties already on record, and chases what is missing before the report is closed out. When the reporter is still on the phone, it asks in the moment, which is the only time that question is cheap to answer. When the report came by email, the follow-up goes back by email, in a thread the next person can read.

What reaches your team is not a task to key in. It is a filed claim: policy matched, parties identified, loss described in your own fields, with the original recording or document attached to it.

Why it matters. Intake quality sets the ceiling on everything downstream. A claim that arrives complete can be triaged, reserved and assigned the same day. A claim that arrives as a half legible form becomes three follow-up calls that nobody has time to make.

A claimant wants to know whether the check has gone out. An employer wants to know what is being paid and until when. A provider wants to know whether a treatment was authorized. These calls are short, they are predictable, and in most claims shops they are the largest single source of inbound volume.

The status agent answers them. It verifies the caller against the record, reads the current state of the claim, and explains it in plain language: what has been paid, what is pending, what has been requested from whom, and what happens next. There is no hold while someone opens the file, because the file is already open.

It knows the limits of its brief. It reports what the record says. It does not interpret coverage, it does not argue a denial, and it does not improvise a date that nothing in the file supports. When a caller asks for any of those, it says so and routes the call to the person who owns the claim, with a summary of what was already asked.

Outbound is the work that quietly does not happen. Wage statements, records requests, authorization follow-ups and welfare checks all sit on a worklist with no deadline attached, and a claim ages while they wait. The outreach agent works that list.

It calls employers for wage and return to work information, providers for records and appointment availability, and injured parties for the detail a claim needs to move. It leaves messages when nobody answers, tries again on a cadence your team sets, and stops when it has what it was sent for.

Every attempt is written to the file: who was reached, when, what was said, what was promised and what arrived. The result is a chase log that reflects reality rather than a reminder that somebody meant to call.

By the time a claim is contested, the file holds hundreds of pages written by people who never expected them to be read together: records, correspondence, forms, invoices, notes. The answer to most questions about that claim is already somewhere in there. Finding it is a research task performed under time pressure.

The assistant does the research. Ask it what is outstanding, whether the mechanism of injury is consistent across the file, what a provider actually said in a report, or what has changed since the last review, and it pulls the relevant passages and answers from them, with a reference to every document and note it used.

It will also tell you what it could not find. An answer the file does not support is not an answer we are willing to generate, and a recommendation is only ever a recommendation: the adjuster reads the source, makes the call, and owns it.